Business · Research · Engineering
Ideas you can examine.
The writing makes my methods open to examination. Start with a business question, follow an experiment, or browse the complete record. Research results, operating proposals and reading selections each retain their own evidence and limits.
The operating advantage
Software gets easier.
Advantage doesn’t.
When software becomes easier to build, the strategic question shifts: which combination of customer access, workflow knowledge, useful data, and organizational capability can turn it into an advantage?
This is the thread connecting the business questions, engineering experiments, research, and reading shelf. The methods are public. Their business applicability still has to be tested.
Go beneath the strategy
Executable notebooks →Engineering →
Systems, emulator correctness, and experiments you can reproduce.
NFC Cannot Fix the Wrong DiacriticResearch →
Romanian language models, synthetic narratives, and careful evaluation.
Merging LoRA Removes a Branch, Not ParametersApplied Papers →
What research findings could change in business and operations.
Deep Research Needs an Evidence Budget44 field notes88 papers mapped3 executable notebooks
Recent work
All notes →Deep Research Needs an Evidence Budget
Deep-research agents need explicit budgets for discovery, inspection, verification, and stopping before their reports can support business decisions.
Merging LoRA Removes a Branch, Not Parameters
A merged LoRA checkpoint removes adapter computation at inference, but it stores a full dense model and preserves only the low-rank origin of the update.
A Test Set Cannot Choose Its Own Threshold
Choosing a classifier cutoff on the final test labels turns evaluation noise into an optimistic result, even when the score order contains no signal.
A Distillation Weight Needs Its Temperature
A knowledge-distillation coefficient is not reproducible unless the temperature, reduction, and valid-token denominator are recorded with it.
An ongoing research thread
Small models.
Connected questions.
TinyFabulist connects controlled story generation, English–Romanian translation, and compact Romanian language models. The research notes follow the data, the experiments, and the limits of what was measured.
Explore the research →Selected work
Start here
When Noisy Text Changes the Winning Model
Romanian diacritic restoration changes winners when clean benchmark text gives way to typos and OCR-like corruption.
Change One Slot, Watch What Else Moves
A paired intervention test separates response to a requested field from leakage into a field that should stay fixed.
Three Emulator Bugs, Three Different Tests
Printer output, audio hardware, and memory bookkeeping needed different tests, and the same discipline about what each fix established.
A Defensive Ollama Client for Local Judge Inference
A corrected native Ollama API example with schema-constrained output, non-streaming responses, timeouts, parsing, and validation.
The reading shelf
A better question
for the next decision.
Systems, technical leadership, and machine learning. A small shelf organized around the problems a book might help you think through.
Explore the books →- Systems & decisionsThinking in Systems
What keeps recreating the problem?
- Technical leadershipScaling People
Which decisions still depend on you being in the room?
- Engineering & maintenanceWorking in Public
Who maintains the software everyone depends on?
Project chronology
- 07 Oct 2026Perplexity Has a Tokenizer in Its Denominatorresearch note
- 05 Oct 2026NFC Cannot Fix the Wrong Diacriticengineering note
- 02 Oct 2026Three Judge Calls Are Still One Itemresearch note
- 30 Sep 2026Count the Raster After Decodingengineering note
- 28 Sep 2026A Ranking Metric Cannot Choose a Thresholdresearch note
- 25 Sep 2026The Perceptron Convergence Bound Has Unitsresearch note